GPT-LAD: Leveraging Large Multimodal Models for Logical Anomaly Detection

As detecting both structural and logical anomalies becomes crucial for robust anomaly detection, recent research has been focusing on logical anomaly detection that can identify both types of anomalies. In actual industrial environments, inspectors are usually supplied with predefined criteria for normal conditions to evaluate whether the produced items are normal or abnormal. However, no existing methods replicate the way humans set and compare against normality criteria to judge normality. To address this gap, we propose a novel framework that mimics human reasoning by defining normality criteria and leveraging GPT-4V’s advanced logical reasoning capabilities. By combining GPT-4V with domain-specific CNN-based models, our approach achieves state-of-the-art or competitive performance on the MVTec LOCO benchmark dataset, offering a more human-like and efficient solution for logical anomaly detection.